Automatic extraction and quantitative analysis of characteristics from complex fractures on rock surfaces via deep learning

Abstract

The detection and evaluation of rock mass joints and fractures are essential in assessing the stability of engineering rock masses and mitigating geological hazards. To address the challenge of intelligent extraction and quantification of fractures, a deep learning-based complex rock fracture segmentation network, termed CRFSegNet, has been developed and combined with multiple feature computation methods. Ablation experiments and multi-model comparisons are conducted on a self-constructed dataset comprising fractures induced by natural processes and blasting. CRFSegNet performs competitively in terms of visualization and evaluation metrics in comparative experiments, with an average intersection-over-union of 83.90 %. The network effectively captures the intricate characteristics of fractures, demonstrating the approach's robustness and competitiveness. Fracture characteristics, such as length-dip, surface fracture rate, and fractal dimension, are obtained based on the segmentation results and the proposed characteristic calculation method. By analyzing the feature acquisition of four images, it is found that the results based on CRFSegNet are basically consistent with the actual situation, which shows that the proposed method is an effective approach for intelligent recognition and feature acquisition.